Deep visual domain adaptation: A survey

difference between one step and multi-step

Class criterion

soft label

soft max

metric learning

（constructed the semantic alignment loss and the separation loss accordingly）

MMD

Architecture criterion

adaptive batch normalization (BN)

weak- related weight



Domain- guided dropout
it mutes non-related neurons for each domain.
（pseudo labels and attribute representation）

Geometric criterion:

This criterion assumes that the relationship of geometric structures can reduce the domain shift

DLID generates in- termediate datasets, starting with all the source data samples and gradually replacing source data with target data

Adversarial-based approaches

Reconstruction-based approaches

Multi-step domain adaptation

Representation-based approaches freeze the previously trained network and use their intermediate representations as input to the new network.